Consistent Instance False Positive Improves Fairness in Face Recognition (2106.05519v1)

Abstract: Demographic bias is a significant challenge in practical face recognition systems. Existing methods heavily rely on accurate demographic annotations. However, such annotations are usually unavailable in real scenarios. Moreover, these methods are typically designed for a specific demographic group and are not general enough. In this paper, we propose a false positive rate penalty loss, which mitigates face recognition bias by increasing the consistency of instance False Positive Rate (FPR). Specifically, we first define the instance FPR as the ratio between the number of the non-target similarities above a unified threshold and the total number of the non-target similarities. The unified threshold is estimated for a given total FPR. Then, an additional penalty term, which is in proportion to the ratio of instance FPR overall FPR, is introduced into the denominator of the softmax-based loss. The larger the instance FPR, the larger the penalty. By such unequal penalties, the instance FPRs are supposed to be consistent. Compared with the previous debiasing methods, our method requires no demographic annotations. Thus, it can mitigate the bias among demographic groups divided by various attributes, and these attributes are not needed to be previously predefined during training. Extensive experimental results on popular benchmarks demonstrate the superiority of our method over state-of-the-art competitors. Code and trained models are available at https://github.com/Tencent/TFace.

Consistent Instance False Positive Improves Fairness in Face Recognition

The paper "Consistent Instance False Positive Improves Fairness in Face Recognition" addresses the persistent issue of demographic bias in face recognition systems. Traditional methods to mitigate this bias often rely on extensive demographic annotations, which are impractical or unavailable in real-world scenarios. Additionally, existing approaches are typically tailored to specific demographic groups, limiting their general applicability. This research introduces a novel method to enhance fairness in these systems without requiring demographic annotations, thus broadening its utility.

Methodology

The core innovation of this paper is the introduction of a false positive rate (FPR) penalty loss. This loss function is designed to reduce bias by promoting the consistency of instance-level FPRs. Unlike prior methods that focus on demographic group-level fairness, this approach targets individual instances, making it adaptable to groups defined by varied attributes without needing explicit annotations during training.

The FPR at the instance level is defined as the ratio of non-target similarities exceeding a unified threshold to the total number of non-target similarities. A penalty proportional to the ratio of instance FPR to the overall FPR is added to the denominator of the softmax-based loss. This unequal penalization encourages consistency across instances, effectively generalizing fairness improvement across demographic groups.

Experimental Results

The efficacy of the proposed method is demonstrated through extensive experiments on prominent facial recognition benchmarks. The results reveal that this approach outperforms state-of-the-art competitors, achieving better fairness metrics and accuracy improvements without relying on demographic annotations.

Notably, the method addresses the larger variance in FPR across demographic groups compared to FNR, which has been acknowledged but underexplored in previous literature. By enhancing the consistency of FPR across identities, the proposed loss function effectively mitigates racial bias without compromising overall recognition performance.

Implications and Future Directions

The implications of this research are significant for the development of fairer face recognition systems, particularly in contexts where demographic annotations are unavailable or demographic groups are not predefined. This loss function can be seamlessly integrated into existing softmax-based architectures, offering a practical solution for the industry.

Looking forward, this work opens several avenues for further exploration. Future studies could investigate alternative penalty functions for inconsistency reduction or explore methodologies to handle potential noise in training data that might be erroneously treated as false positives. Additionally, this framework could be extended to other domains in AI where fairness and bias are critical concerns.

In conclusion, this paper presents a robust solution to a challenging problem in face recognition, enhancing both fairness and performance metrics without the need for extensive demographic data. This advancement is a crucial step toward more equitable AI systems in various applications.